Obaseki
Resume
AI agents

Tektradoor trading agent

An AI agent for crypto futures, rebuilt in public with every strategy pre-registered before it is tested.

The test rig is open source.

Stack
PythonBacktestingOpenTimestamps
Status
Rebuilding in public

The problem

Tektradoor’s first trading bot didn’t work. Too many things went wrong to patch, so we shut it down and started over, in public.

The hardest problem in automated trading isn’t writing a strategy. It’s knowing whether a strategy really works, or whether it only looks good because someone kept adjusting it until the backtest did. A result you can tune after the fact proves nothing.

What I built

Before writing a new agent, I built the test system it has to pass: tektradoor-judge, a pre-registered test rig.

  • Write it down first. Every candidate strategy is written down before it runs: its rules, its random seed and the exact code version.
  • Timestamp it in public. The file’s fingerprint (sha256) is posted publicly and stamped with OpenTimestamps, which anchors it in a Bitcoin block. Anyone can check that the rules existed before the test.
  • Test on unseen data. The strategy then runs on data it hasn’t seen, with trading costs, funding and a buy-and-hold benchmark applied the same way to every candidate.
  • Refuse to cheat. The judge won’t run if the registered file or the judge’s own code changed after registration. Results are committed and never overwritten, and the final test can run only once.
  1. Write the strategy down rules, seed, code version
  2. Fingerprint and timestamp sha256, posted and stamped
  3. Test on unseen data costs and benchmark applied equally
  4. Final run, once results committed, never overwritten
Parts I built

Key decisions

  • The test system comes before the agent. A new strategy is only worth building if there’s a way to judge it that can’t be tuned afterwards.
  • Every result gets posted, losses included.
  • The rig states its own limits. Whoever holds a repository can rewrite its history, so the git checks only catch mistakes. The real proof is outside the repo: the public fingerprint posts, GitHub’s push times and the Bitcoin-anchored timestamps.

Results

  • The test rig is open source.
  • The agent is being rebuilt in public on top of it.